A Hermes-like, self-evolving AI agent harness — running entirely inside Chrome, no local server required.
NanoHarness is a pure Chrome-extension AI agent harness: an agent loop, browser automation, a workspace filesystem, a skill system, memory & self-evolution — and it can expose itself as an MCP server for other agents to call. Everything runs inside the browser; no local server needed.
NanoHarness doesn't just call tools — like Hermes, it learns who you are and gets better at your tasks:
- Two-channel memory: remembers who you are (preferences, habits) and environment conventions/pitfalls across sessions, so you never repeat yourself.
- Skill self-evolution: after finishing a reusable task, it proactively saves it as a skill, then reuses it later. Skills accumulate — the agent gets better at your specific work.
- Automatic context compression: long conversations are auto-summarized so the context window never overflows — long tasks don't "lose memory".
- Asks only when it matters: decides low-stakes choices itself; pauses with options only when there's a real trade-off.
Drive the browser over CDP: open/switch tabs, read page content, click, fill forms, screenshot — 20+ browser tools. The agent can genuinely see and operate your browser.
Record your clicks, typing, scrolling and hovering into a reusable skill. Next time, just say it in chat and the whole flow replays automatically.
Expose the extension as a standard MCP server (Streamable HTTP). Claude Desktop, Cursor, or any MCP client can call your browser, workspace files and skills.
Add/remove/edit skills and configure remote MCP servers (HTTP/SSE) in Settings. Supports OpenAI-compatible and Anthropic protocols — DeepSeek, OpenAI, Claude, Moonshot, GLM, Qwen, and more.
Both GIFs walk through the same flow: ask "Who are you?" → "Open bing" → browse every Settings tab (with token usage).
English:
中文 / Chinese:
- Clone:
git clone https://github.com/anyforge/nanoharness.git - Open Chrome and go to
chrome://extensions - Turn on Developer mode (top-right)
- Click Load unpacked and select this repository's root folder
- Click the NanoHarness toolbar icon to open the side panel
- Side panel → ⚙ Settings → Model, fill Base URL / API Key / model (presets fill most of it), save.
- Back in chat, just say what you want, e.g. "open this page and extract the key points".
- To turn your own actions into a skill: hit Record → use the page normally → stop → save as a skill.
NanoHarness is an automation tool. It genuinely operates the browser and calls external LLMs and MCP services. Do not use it where you don't understand the consequences — e.g. around accounts, payments, private data, or production systems. You are responsible for whatever it does on your behalf. The software is provided "as is", without warranty of any kind; the author assumes no liability for any direct or indirect loss.